Enhancing image resolution
Methods, machines, and machine-readable media for enhancing image resolution are described. In one aspect, a respective motion map is computed for each pairing of a reference image and a respective image neighboring the reference image in a sequence of base images. Each motion map includes a set of motion vectors mapping reference image pixels to respective neighboring image pixels. Respective regions of a target image are assigned to motion classes based on the computed motion maps. The target image has a target resolution level and the base images have a base resolution level equal to or lower than the target resolution level. Pixel values for the target image are computed based on corresponding pixel value contributions from the motion-compensated neighboring base images selected in accordance with the motion classes assigned to the target image regions.
Latest Hewlett Packard Patents:
The resolution of an image is determined by the physical characteristics of the sensor that is used to capture the image. In some resolution enhancement approaches, the spatial dimension of an image is increased by interpolating between pixel values of a single image to generate an enhanced image. Both linear and non-linear interpolation techniques have been applied to a single image to produce an image with a higher spatial dimension than the original image. In these resolution enhancement techniques, however, the overall information content in the enhanced image is the same as the original image: the pixel number is increased, but the resolution is not improved. In other approaches, image resolution is increased by obtaining multiple displaced images of the scene and combining the information from the multiple images into an enhanced image. The process of reconstructing an image from several displaced images of the same scene often is referred to as “super resolution”. In these resolution enhancement techniques, the overall information content in the enhanced image may be greater than the information content in the individual original images.
In some super resolution methods, one of the lower resolution images in a sequence of lower resolution images is designated as a reference image. The coordinate frame of the reference image is up-projected to the higher resolution level to define the coordinate frame of the higher resolution image. Motion vectors between the reference image and each of a selected number of lower-resolution images neighboring the reference image in the multi-image sequence are estimated. The motion vectors then are up-projected to the coordinate frame of the higher resolution image. The up-projected motion vectors are used to re-map the neighboring images to the coordinate frame of the higher resolution image. Next, the coordinate frame of the higher resolution image is populated with pixel values that are computed based on combinations of pixel values from the re-mapped lower resolution images. In some approaches, a non-linear interpolation technique is used to reconstruct the higher resolution image from the re-mapped pixel values.
SUMMARYThe invention features methods, machines, and machine-readable media for enhancing image resolution.
In one aspect of the invention, a respective motion map is computed for each pairing of a reference image and a respective image neighboring the reference image in a sequence of base images. Each motion map comprises a set of motion vectors mapping reference image pixels to respective neighboring image pixels. Respective regions of a target image are assigned to motion classes based on the computed motion maps. The target image has a target resolution level and the base images have a base resolution level equal to or lower than the target resolution level. Pixel values for the target image are computed based on corresponding pixel value contributions from the base images selected in accordance with the motion classes assigned to the target image regions.
Other features and advantages of the invention will become apparent from the following description, including the drawings and the claims.
In the following description, like reference numbers are used to identify like elements. Furthermore, the drawings are intended to illustrate major features of exemplary embodiments in a diagrammatic manner. The drawings are not intended to depict every feature of actual embodiments nor relative dimensions of the depicted elements, and are not drawn to scale.
The image processing embodiments described below incorporate a dynamic approach to enhance the spatial resolution of an image sequence that allows different regions of the scene captured in the images to be treated differently. In this way, these embodiments are able to avoid artifacts that otherwise might result from treating all regions of the scene in the same way during the resolution enhancement process. In addition, these embodiments are able to dynamically tailor the image resolution enhance process in an intelligent way. In particular, these embodiments deploy image processing resources to different regions of an enhanced resolution image at varying computational intensity levels to achieve high quality resolution enhancement results in an accurate and efficient way.
In general, the modules 14-20 of system 10 are not limited to any particular hardware or software configuration, but rather they may be implemented in any computing or processing environment, including in digital electronic circuitry or in computer hardware, firmware, device driver, or software. For example, in some implementations, these modules 14-20 may be embedded in the hardware of any one of a wide variety of digital and analog electronic devices, including desktop and workstation computers, digital still image cameras, digital video cameras, printers, scanners, and portable electronic devices (e.g., mobile phones, laptop and notebook computers, and personal digital assistants).
The base image sequence 12 may correspond to an original base image sequence that was captured by an image sensor (e.g., a video image sequence or a still image sequence) or a processed version of such an original base image sequence. For example, the base image sequence 12 may consist of a sampling of the images selected from an original base image sequence that was captured by an image sensor or a compressed or reduced-resolution version of an original base image sequence that was captured by an image sensor. In order to achieve spatial resolution enhancement of the base image sequence 12, at least some of the base images correspond to displaced images of the same scene so that different samplings of the scene can be combined into an enhanced resolution target image.
Each target image 22 is produced from pixel value contributions from a selected set of the base images 12, including one that is designated the “reference image” and one or more base images that neighbor the reference image in the sequence. As used herein, the term “neighboring base images” refers to base images within a prescribed number of base images of each other in a base image sequence without regard to the temporal ordering of the neighboring base images in terms of capture time. In addition, the term “successive base images” refers to adjacent base images in a base image sequence that may be ordered chronologically or reverse-chronologically in terms of capture time. The number of neighboring images used to compute a target image, and the relative positions of the neighboring images in the sequence, are implementation-specific parameters. In some implementations, three successive neighboring base images on either side of the reference image in the base image sequence are processed with each reference image, for a total of seven base images that are processed for each target image.
Referring to
Ux(x,y)=ax0+ax1x+ax2y (1)
Uy(x,y)=ay0+ay1x+ay2y (2)
wherein Ux(x,y) and Uy(x,y) are the x and y components of a velocity motion vector at point (x,y), respectively, and the ak's are the affine motion parameters.
Examples of an affine motion model are illustrated in
In some embodiments, the motion maps of image pairs are represented as vector fields in the coordinate system of the reference image, which defines the coordinate system of the target image to be enhanced. A vector field U(P) the reference image Ir(P), and the neighboring image It(P) (e.g., one of the images preceding or succeeding the image to be enhanced in a image sequence), satisfy:
Ir(P)=It(P−U(P)) (3)
where P=P(x,y) represents pixel coordinates. Therefore, each of the neighboring images can be warped to the coordinate frame of the corresponding reference image using equation (3) to create:
Itw(P)=It(P−U(P)) (4)
where Itw(P) is the warped neighboring image.
In a typical image sequence of base images, if the motion vectors are computed correctly, the warped neighboring image should look very similar to the corresponding reference image. In the case of video sequences, the reference and neighboring images are captured at two different times. As a result, the pixel motion between images is due to both camera motion and also the motion of scene points moving independently. The motion of pixels is therefore unconstrained (non-parametric). Accordingly, in some embodiments, motion estimation module 14 computes movements of individual pixels or groups of pixels from a given base image to a neighboring base image based on a non-parametric optical flow model (or dense motion model). The motion estimates may be computed for one or both of forward and backwards transitions in time (i.e., from an earlier-captured base image to a later-captured base image, or from a later-captured base image to an earlier-captured base image).
In some embodiments, motion is modeled as a smoothly varying flow field, and motion analysis exploits local smoothness of optical flow fields. In this optical flow model, the movements are represented by velocity vectors (dx/dt, dy/dt) that describe how quickly a pixel (or a group of pixels) is moving across an image, and the direction of pixel movement. The optical flow model represents a projection of three-dimensional object motion onto the image sensor's two-dimensional image plane. Any one of a wide variety of optical flow computation methods may be used by the motion estimation module 14 to compute motion vectors. In some implementations, a multi-scale coarse-to-fine algorithm based on a gradient approach may be used to compute the optical flow motion vectors.
In some of these embodiments, the reference and neighboring image pairs are represented by Laplacian or Gaussian multi-resolution pyramids. In this way, these embodiments are able to accommodate a wide range of displacements, while avoiding excessive use of computational resources and generation of false matches. In particular, using a multi-resolution pyramid approach allows large displacements to be computed at low spatial resolution. Images at higher spatial resolution are used to improve the accuracy of displacement estimation by incrementally estimating finer displacements. Another advantage of using image pyramids is the reduction of false matches, which is caused mainly by the mismatches at higher resolutions under large motion. Motion estimation in a multi-resolution framework helps to eliminate problems of this type, since larger displacements are computed using images of lower spatial resolution, where they become small displacements due to sub-sampling.
In these embodiments, motion estimation module 14 uses a pyramid-based hierarchical image alignment technique to align two input images (i.e., a neighboring image and a corresponding reference image). A Laplacian or Gaussian pyramid is constructed from each of the two input images, and motion parameters are estimated in a coarse-to-fine manner. Within each pyramid level the sum of squared differences (SSD) measure integrated over regions of interest (which is initially the entire image region) is used as a match measure:
where I is the Laplacian or Gaussian filtered image intensity. The sum is computed over all the points P within the region and is used to denote the SSD error of the entire motion field within that region. The motion field is modeled by a set of global parameters (e.g., plane parameters) and local parameters (e.g., optical flow) as described above.
Numerical methods such as Gauss-Newton minimization is applied to the objective function described in equation (5) in order to estimate the unknown motion parameters and the resulting motion field. Starting with some initial values (typically zero), the hierarchical estimation algorithm iteratively refines the parameters in order to minimize the SSD error described in equation (5) from coarse to fine resolutions. After each motion estimation step, the current set of parameters is used to warp the neighboring image to the coordinate frame of the reference image, as described in equation (4), in order to reduce the residual displacement between the images.
The optical flow at each pixel is assumed to be locally constant in a small window around that pixel. The flow for a pixel is estimated by using all the pixels in its window. This process is repeated for each pixel and results in a smoothly varying flow field. In some implementations, dense optical flow is estimated using five windows 30, 32, 34, 36, 38, on and off-centered around each pixel under examination, as illustrated in
Referring back to
where σref2 and σremapped2 are the respective reference and remapped neighboring image variances within the correlation window; σN2=σ2/(μ2+c) is the mean normalized variance with μ being the mean and c a stabilizing constant to handle close-to-zero mean values; Ω, ΩN and κ are thresholding parameters, and N is the number of pixels in the correlation window.
In some implementations, the total alignment quality is determined by computing the geometric mean of the quality measure for each of the color (e.g., Red, Green, and Blue) spectral bands of the base images 12. The alignment quality measures for each pairing of the reference image and a respective neighboring image are contained in respective alignment accuracy maps.
Referring back to
The motion segmentation module 16 then segments pixels in the down-sampled motion magnitude maps 48 into motion classes (block 50). Motion segmentation module 16 may classify pixel regions in each down-sampled motion magnitude map into a respective set of motion classes (e.g., a high motion class region 58, intermediate motion class regions 54, 56, and a low motion class region 52) using any type of classification or segmentation method. For example, in some implementations, motion vectors in each motion magnitude map are segmented in accordance with a k-means clustering method. In these implementations, either the number of clusters or a set of clusters representing an initial partition between motion magnitudes in a given motion magnitude map may be pre-determined. The partition is refined iteratively by assigning pixels to each partition and re-computing the center of each cluster. The segmentation method iterates between the following steps:
1. Compute cluster centroids and use them as new cluster seeds; and
2. Assign each object to the nearest seed.
In some implementations, the final partition corresponds to a respective set of motion magnitude clusters 52, 54, 56, 58 for each coarse motion magnitude map in which the total distance between pixels (or pixel groups) and the centers of their respective clusters is minimized, while the distances between clusters are maximized.
The motion segmentation module 16 iteratively segments pixel motion magnitude maps from coarser resolution levels up to the original base image resolution level (block 60). The segmentation results from the previous resolution are used as the starting point for the same segmentation process applied to the next higher resolution level.
The separate motion class segmentation maps that are computed for each pairing of the reference image and a respective neighboring image are merged into a unified motion class segmentation map for the reference image. In some implementations, the motion segmentation module 16 assigns a given reference image pixel to the low motion class in the unified motion class segmentation map when the given pixel is assigned to the low motion class in all of the separate motion class segmentation maps. The motion segmentation module 16 assigns a given reference image pixel to the high motion class in the unified motion class segmentation map when the given pixel is assigned to the high motion class in any of the separate motion class segmentation maps. As explained above, some implementations include an intermediate motion class, where motion vectors that are assigned to the intermediate motion class have magnitudes higher than motion vectors assigned to the low motion class and lower than motion vectors assigned to the high motion class. In these implementations, the motion segmentation module 16 assigns a given reference image pixel to the intermediate motion class in the unified motion class segmentation map when the given pixel is unassigned to the high motion class in any of the separate motion class segmentation maps and is unassigned to the low motion class in all of the separate motion class segmentation maps.
Referring back to
Based on the up-projected motion maps, the adaptive synthesis module 20 re-maps the neighboring images to the coordinate frame of the target image (block 64). In some implementations, the neighboring images are up-projected to the target image resolution level and the up-projected neighboring images are re-mapped to the target image coordinate frame using the up-projected motion maps in accordance with application of equation (4) at the target image resolution level.
The adaptive synthesis module 20 applies a threshold to the up-projected alignment accuracy maps to produce respective synthesis maps for each of the neighboring images (block 66). The synthesis maps are used by the adaptive synthesis module 20 to classify motion vectors in each up-projected motion map into valid and invalid motion vector classes. In the illustrated embodiment, the threshold that is applied to the up-projected alignment accuracy maps is set to a level that ensures that the neighboring image pixels are sufficiently aligned with respect to the corresponding reference image pixels that they contain relevant information for reconstructing the target image. In some implementations, the alignment measures in the alignment accuracy maps are normalized to values in the range of ±1 and the threshold is set to approximately 0.8. Pixels in the up-projected motion maps with alignment accuracy measures above the threshold are classified as valid motion pixels, whereas pixels with alignment accuracy measures below the threshold are classified as invalid motion pixels.
Referring to
For pixels in the neighboring images that are identified as having invalid motion vectors according to the synthesis maps, their contribution to the target image is zero. For target image pixels whose corresponding neighboring image pixels are all identified as having invalid motion vectors (block 70), the adaptive synthesis module 20 computes target image pixel values by interpolating up-projected reference image pixel values (block 72). In some implementations, the adaptive synthesis module 20 uses bi-cubic interpolation to compute values of target image pixels associated with invalid motion vectors in all neighboring images.
For target image pixels assigned to the high motion class with at least one neighboring image pixel with valid motion vectors (blocks 70, 74), the adaptive synthesis module 20 computes pixel values by merging contributions from the reference image and the two relevant nearest re-mapped neighboring images (block 76). In some implementations, the high motion class target image pixels are computed by a pixel-wise weighted combination given by equation (8):
where ITarget is the synthesized target image, ωT
For target image pixels assigned to the intermediate and low motion classes (block 70, 74), the adaptive synthesis module 20 computes measures of local texture richness (block 78). Texture descriptors can be statistical, structural, or syntactic. In some implementations, a statistical descriptor is used. In these implementations, for a small local region around each pixel in the intermediate and low motion class regions, the adaptive synthesis module 20 computes the standard deviation of the Laplacian image, the skewness of the gradient value distribution, and the edge frequency as measures of local texture content. In some implementations, both the gradient image and the Laplacian image are computed during the motion estimation process (block 24;
The adaptive synthesis module 20 segments intermediate and low motion class pixels into high and low texture region classes based on the computed local texture richness measures (block 80). In some implementations, if the computed texture richness measure is below an empirically-determined threshold value, the adaptive synthesis module 20 segments the pixels into the low texture region class; otherwise the pixels are segmented into the high texture region class.
For intermediate and low motion class pixels assigned to the low texture region class in the reference image and all the corresponding neighboring images (block 82), the adaptive synthesis module 20 computes corresponding target image pixel values by interpolating up-projected reference image values (block 84). For intermediate and low motion class pixels assigned to the high texture region class in the reference image or any of the corresponding neighboring images (block 82), the adaptive synthesis module 20 computes target image pixel values by merging contributions from the up-projected reference image and the relevant re-mapped neighboring images in accordance with equation (9) (block 86):
where ωTexture
The above-described resolution enhancement embodiments may be applied to one or all of the luminance and chrominance components of the base images 12. In some embodiments, the resulting resolution-enhanced target images 22 may be subjected to one or more post-processing methods, including color re-mapping, sharpening, and de-scintillation methods.
Other embodiments are within the scope of the claims.
The systems and methods described herein are not limited to any particular hardware or software configuration, but rather they may be implemented in any computing or processing environment, including in digital electronic circuitry or in computer hardware, firmware, or software. In general, the systems may be implemented, in part, in a computer process product tangibly embodied in a machine-readable storage device for execution by a computer processor. In some embodiments, these systems preferably are implemented in a high level procedural or object oriented processing language; however, the algorithms may be implemented in assembly or machine language, if desired. In any case, the processing language may be a compiled or interpreted language. The methods described herein may be performed by a computer processor executing instructions organized, for example, into process modules to carry out these methods by operating on input data and generating output. Suitable processors include, for example, both general and special purpose microprocessors. Generally, a processor receives instructions and data from a read-only memory and/or a random access memory. Storage devices suitable for tangibly embodying computer process instructions include all forms of non-volatile memory, including, for example, semiconductor memory devices, such as EPROM, EEPROM, and flash memory devices; magnetic disks such as internal hard disks and removable disks; magneto-optical disks; and CD-ROM. Any of the foregoing technologies may be supplemented by or incorporated in specially designed ASICs (application-specific integrated circuits).
Claims
1. A machine-implemented image processing method, comprising:
- computing a respective motion map for each pairing of a reference image and a respective image neighboring the reference image in a sequence of base images, each motion map comprising a set of motion vectors mapping reference image pixels to respective neighboring image pixels;
- assigning respective regions of a target image to motion classes based on the computed motion maps, the target image having a target resolution level and the base images having a base resolution level equal to or lower than the target resolution level; and
- computing pixel values for the target image based on corresponding pixel value contributions from the base images selected in accordance with the motion classes assigned to the target image regions.
2. The method of claim 1, wherein computing motion maps comprises generating for each image pair respective dense motion vectors describing motion at pixel locations with respective sets of parameters in a motion parameter space at sub-pixel accuracy.
3. The method of claim 1, wherein assigning regions of the target image to motion classes comprises assigning pixels of the reference image to respective motion classes and up-projecting the motion class assignments to pixels of the target image.
4. The method of claim 3, wherein assigning reference image pixels to motion classes comprises computing motion magnitude maps from each motion map, down-sampling the computed motion magnitude maps to a pyramid of motion magnitude maps at respective resolution levels lower than the base resolution level, and segmenting pixels in the pyramid of down-sampled motion magnitude maps into motion classes.
5. The method of claim 4, wherein assigning reference image pixels to motion classes further comprises iteratively segmenting pixels in the pyramid of motion magnitude maps from a coarsest resolution level up to the base resolution level, wherein segmentation from each coarser resolution level is up-sampled to initialize the segmentation at a finer resolution level, and refined by further segmentation at the finer resolution level.
6. The method of claim 3, wherein assigning reference image pixels to motion classes comprises generating a separate motion class segmentation map for each pairing of the reference image and a respective neighboring image and merging the separate motion class segmentation maps into a unified motion class segmentation map for the reference image.
7. The method of claim 6, wherein reference image pixels are respectively assigned to a motion class selected from a motion class set including a high motion class and a low motion class, and motion vectors assigned to the high motion class have higher magnitudes than motion vectors assigned to the low motion class.
8. The method of claim 7, wherein merging the separate motion class segmentation maps comprises
- assigning a given reference image pixel to the low motion class in the unified motion class segmentation map when the given pixel is assigned to the low motion class in all of the separate motion class segmentation maps, and
- assigning a given reference image pixel to the high motion class in the unified motion class segmentation map when the given pixel is assigned to the high motion class in any of the separate motion class segmentation maps.
9. The method of claim 8, wherein the motion class set further includes an intermediate motion class, and motion vectors assigned to the intermediate motion class have magnitudes higher than motion vectors assigned to the low motion class and lower than motion vectors assigned to the high motion class.
10. The method of claim 9, wherein merging the separate motion class segmentation maps comprises assigning a given reference image pixel to the intermediate motion class in the unified motion class segmentation map when the given pixel is unassigned to the high motion class in any of the separate motion class segmentation maps and is unassigned to the low motion class in all of the separate motion class segmentation maps.
11. The method of claim 1, further comprising computing an alignment accuracy map for each pairing of the reference image and a respective neighboring image based on the computed motion maps.
12. The method of claim 11, wherein computing alignment accuracy maps comprises re-mapping neighboring images to a coordinate frame of the reference image using respective motion maps, and computing correlation measures between pixels of the reference image and pixels of each of the motion-compensated neighboring images.
13. The method of claim 11, further comprising up-projecting the computed alignment accuracy maps from the base image resolution level to the target image resolution level.
14. The method of claim 13, further comprising up-projecting the motion maps from the base image resolution level to the target image resolution level, and classifying motion vectors in each up-projected motion map into valid and invalid motion vector classes based on the up-projected alignment accuracy maps.
15. The method of claim 14, wherein values of target image pixels with corresponding pixels in all neighboring images being associated with motion vectors in the invalid motion vector class are computed by interpolating up-projected pixel values of the reference image.
16. The method of claim 1, further comprising up-projecting the motion maps from the base image resolution level to the target image resolution level.
17. The method of claim 16, further comprising re-mapping the neighboring images to the reference frame of the up-projected reference image using the respective up-projected motion maps.
18. The method of claim 1, wherein regions of the target image are respectively assigned to a motion class selected from a motion class set including a high motion class and a low motion class, and motion vectors of regions assigned to the high motion class have higher magnitudes than motion vectors of regions assigned to the low motion class.
19. The method of claim 18, wherein computing target image pixel values in regions assigned to the high motion class comprises computing a pixel-wise combination of pixel value contributions from the up-projected reference image and the re-mapped neighboring images weighted based on pixel-wise measures of alignment accuracy between the reference image and the corresponding neighboring images.
20. The method of claim 19, wherein pixel value contributions from the re-mapped neighboring images are additionally weighted based on measures of temporal distance between the reference image and the corresponding neighboring images.
21. The method of claim 18, wherein computing target image pixel values in regions assigned to the low motion class comprises classifying low motion class reference image pixels and their corresponding pixels in the neighboring images based on measures of local texture richness.
22. The method of claim 21, wherein low motion class reference image pixels and their corresponding pixels in the neighboring images are quantitatively evaluated for local texture richness, and are classified into a texture class selected from the texture class set including a high texture region class and a low texture region class, and pixels assigned to the high texture region class have higher local texture measures than pixels assigned to the low texture region class.
23. The method of claim 22, wherein values of target image pixels classified into the low texture region class in the reference image and all of the respective neighboring images, are computed by interpolating up-projected pixel values of the reference image.
24. The method of claim 22, wherein a value of a given target image pixel classified into the high texture region class in the reference image or any respective neighboring images is computed based on a pixel value contribution from the up-projected reference image, and a pixel value contribution from a given re-mapped neighboring image weighted based on a measure of local texture richness computed for the given pixel, a measure of motion estimation accuracy computed for the given pixel, and a measure of temporal distance of the neighboring image from the reference image.
25. The method of claim 18, wherein values of target image pixels are computed based on pixel value contributions from a number of base images neighboring the reference image, the number of neighboring base images being different for different motion classes.
26. The method of claim 25, wherein values of target image pixels assigned to the high motion class are computed based on pixel value contributions from a fewer number of neighboring base images than the values of target image pixels assigned to the low motion class.
27. The method of claim 26, wherein the motion class set further includes an intermediate motion class, the motion vectors of regions assigned to the intermediate motion class have lower magnitudes than motion vectors of regions assigned to the high motion class and higher magnitudes than motion vectors of regions assigned to the low motion class, and values of target image pixels assigned to the intermediate motion class are computed based on pixel value contributions from a fewer number of neighboring base images than the values of target image pixels assigned to the low motion class but a higher number of neighboring base images than the values of target image pixels assigned to the high motion class.
28. An image processing machine, comprising:
- a computer-readable memory storing computer process instructions; and
- a computer processor operable to execute the computer process instructions and perform operations comprising
- computing a respective motion map for each pairing of a reference image and a respective image neighboring the reference image in a sequence of base images, each motion map comprising a set of motion vectors mapping reference image pixels to respective neighboring image pixels at sub-pixel accuracy,
- assigning respective regions of a target image to motion classes based on the computed motion maps, the target image having a target resolution level and the base images having a base resolution level equal to or lower than the target resolution level, and
- computing pixel values for the target image based on corresponding pixel value contributions from the re-mapped base images selected in accordance with the motion classes assigned to the target image regions.
29. The machine of claim 28, wherein the computer processor is operable to assign regions of the target image to motion classes by assigning pixels of the reference image to respective motion classes and up-projecting the motion class assignments to pixels of the target image.
30. The machine of claim 29, wherein the computer processor is operable to assign reference image pixels to motion classes by computing motion magnitude maps from each motion map, down-sampling the computed motion magnitude maps to a pyramid of motion magnitude maps at respective resolution levels lower than the base resolution level, and segmenting pixels in the pyramid of down-sampled motion magnitude maps into motion classes.
31. The machine of claim 29, wherein the computer processor is operable to assign reference image pixels to motion classes by generating a separate motion class segmentation map for each pairing of the reference image and a respective neighboring image and merging the separate motion class segmentation maps into a unified motion class segmentation map for the reference image.
32. The machine of claim 28, wherein the computer processor is operable to compute an alignment accuracy map for each pairing of the reference image and a respective neighboring image based on the computed motion maps.
33. The machine of claim 32, wherein the computer processor is operable to compute alignment accuracy maps by re-mapped neighboring images to a coordinate frame of the reference image using respective motion maps, and computing correlation measures between pixels of the reference image and pixels of each of the motion-compensated neighboring images.
34. The machine of claim 33, wherein the computer processor is operable to up-project the computed alignment accuracy maps from the base image resolution level to the target image resolution level.
35. The machine of claim 34, wherein the computer processor is operable to up-project the motion maps from the base image resolution level to the target image resolution level, and classify motion vectors in each up-projected motion map into valid and invalid motion vector classes based on the respective up-projected alignment accuracy maps.
36. The machine of claim 35, wherein values of target image pixels with corresponding pixels in all neighboring images being associated with motion vectors in the invalid motion vector class are computed by interpolating up-projected pixel values of the reference image.
37. The machine of claim 28, wherein the computer processor is operable to up-project the motion maps from the base image resolution level to the target image resolution level, and re-map the neighboring images to the reference frame of the up-projected reference image using the respective up-projected motion maps.
38. The machine of claim 28, wherein:
- regions of the target image are respectively assigned to a motion class selected from a motion class set including a high motion class and a low motion class, and motion vectors assigned to the high motion class have higher magnitudes than motion vectors assigned to the low motion class; and
- the computer processor is operable to compute target image pixel values in regions assigned to the high motion class by computing a pixel-wise combination of pixel value contributions from the up-projected reference image and the respective re-mapped neighboring images weighted based on pixel-wise measures of alignment accuracy between the reference image and the corresponding neighboring images.
39. The machine of claim 38, wherein pixel value contributions from the re-mapped neighboring images are additionally weighted based on measures of temporal distance between the reference image and the corresponding neighboring images.
40. The machine of claim 38, wherein the computer processor is operable to compute target image pixel values in regions assigned to the low motion class based on classification of low motion class reference image pixels and corresponding pixels in the neighboring images based on measures of local texture richness.
41. The machine of claim 28, wherein values of target image pixels are computed based on pixel value contributions from a number of re-mapped base images neighboring the reference image, the number of neighboring base images being different for different motion classes.
42. A non transitory Computer readable medium storing computer-readable instructions operable to cause a computer to perform operations comprising:
- computing a respective motion map for each pairing of a reference image and a respective image neighboring the reference image in a sequence of base images, each motion map comprising a set of motion vectors mapping reference image pixels to respective neighboring image pixels at sub-pixel accuracy;
- assigning respective regions of a target image to motion classes based on the computed motion maps, the target image having a target resolution level and the base images having a base resolution level equal to or lower than the target resolution level; and
- computing pixel values for the target image based on corresponding pixel value contributions from the re-mapped base images selected in accordance with the motion classes assigned to the target image regions.
43. The computer-readable medium of claim 42, wherein the computer-readable instructions are operable to cause the computer to assign regions of the target image to motion classes by assigning pixels of the reference image to respective motion classes and up-projecting the motion class assignments to pixels of the target image.
44. The computer-readable medium of claim 43, wherein the computer-readable instructions are operable to cause the computer to assign reference image pixels to motion classes by computing motion magnitude maps from each motion map, down-sampling the computed motion magnitude maps to a pyramid of motion magnitude maps at respective resolution levels lower than the base resolution level, and segmenting pixels in the pyramid of down-sampled motion magnitude maps into motion classes.
45. The computer-readable medium of claim 43, wherein the computer-readable instructions are operable to cause the computer to assign reference image pixels to motion classes by generating a separate motion class segmentation map for each pairing of the reference image and a respective neighboring image and merging the separate motion class segmentation maps into a unified motion class segmentation map for the reference image.
46. The computer-readable medium of claim 42, wherein the computer-readable instructions are operable to cause the computer to compute an alignment accuracy map for each pairing of the reference image and a respective neighboring image based on the computed motion maps.
47. The computer-readable medium of claim 46, wherein the computer-readable instructions are operable to cause the computer to compute alignment accuracy maps by re-mapping neighboring images to a coordinate frame of the reference image using respective motion maps, and computing correlation measures between pixels of the reference image and pixels of each of the motion-compensated neighboring images.
48. The computer-readable medium of claim 47, wherein the computer-readable instructions are operable to cause the computer to up-project the computed alignment accuracy maps from the base image resolution level to the target image resolution level.
49. The computer-readable medium of claim 48, wherein the computer-readable instructions are operable to cause the computer to up-project the motion maps from the base image resolution level to the target image resolution level, and classify motion vectors in each up-projected motion map into valid and invalid motion vector classes based on the respective up-projected alignment accuracy maps.
50. The computer-readable medium of claim 49, wherein values of target image pixels with corresponding pixels in all neighboring images being associated with motion vectors in the invalid motion vector class are computed by interpolating up-projected pixel values of the reference image.
51. The computer-readable medium of claim 42, wherein the computer-readable instructions are operable to cause the computer to up-project the motion maps from the base image resolution level to the target image resolution level, and re-map the neighboring images to the reference frame of the up-projected reference image using the respective up-projected motion maps.
52. The computer-readable medium of claim 42, wherein:
- regions of the target image are respectively assigned to a motion class selected from a motion class set including a high motion class and a low motion class, and motion vectors assigned to the high motion class have higher magnitudes than motion vectors assigned to the low motion class; and
- the computer-readable instructions are operable to cause the computer to compute target image pixel values in regions assigned to the high motion class by computing a pixel-wise combination of pixel value contributions from the re-mapped neighboring images weighted based on pixel-wise measures of alignment accuracy between the reference image and the corresponding neighboring images.
53. The computer-readable medium of claim 52, wherein pixel value contributions from the re-mapped neighboring images are additionally weighted based on measures of temporal distance between the reference image and the corresponding neighboring images.
54. The computer-readable medium of claim 52, wherein the computer-readable instructions are operable to cause the computer to compute target image pixel values in regions assigned to the low motion class based on classification of low motion class reference image pixels and corresponding pixels in the neighboring images based on measures of local texture richness.
55. The computer-readable medium of claim 42, wherein values of target image pixels are computed based on pixel value contributions from a number of re-mapped base images neighboring the reference image, the number of re-mapped neighboring base images being different for different motion classes.
5053875 | October 1, 1991 | Ishii et al. |
5557684 | September 17, 1996 | Wang et al. |
5719594 | February 17, 1998 | Potu |
5973733 | October 26, 1999 | Gove |
6249613 | June 19, 2001 | Crinon et al. |
6269175 | July 31, 2001 | Hanna et al. |
6304682 | October 16, 2001 | Patti |
6307560 | October 23, 2001 | Kondo et al. |
6385245 | May 7, 2002 | De Haan et al. |
6393163 | May 21, 2002 | Burt |
6466618 | October 15, 2002 | Messing et al. |
6476873 | November 5, 2002 | Maeng |
6570612 | May 27, 2003 | Saund et al. |
6587599 | July 1, 2003 | Huang |
6625333 | September 23, 2003 | Wang et al. |
6628805 | September 30, 2003 | Hansen et al. |
6718073 | April 6, 2004 | Kondo et al. |
6782132 | August 24, 2004 | Fogg |
6798834 | September 28, 2004 | Murakami et al. |
6983080 | January 3, 2006 | Wenstrand et al. |
7038729 | May 2, 2006 | Kondo et al. |
7043058 | May 9, 2006 | Cornog et al. |
7088773 | August 8, 2006 | Paniconi et al. |
7119837 | October 10, 2006 | Soupliotis et al. |
7142600 | November 28, 2006 | Schonfeld et al. |
7729558 | June 1, 2010 | Kondo et al. |
20010048763 | December 6, 2001 | Takatsuka et al. |
20020118761 | August 29, 2002 | Lee |
20030012457 | January 16, 2003 | Solecki |
20030016884 | January 23, 2003 | Altunbasak et al. |
20030031382 | February 13, 2003 | Broekaart |
20030044087 | March 6, 2003 | Kondo et al. |
20030053692 | March 20, 2003 | Hong et al. |
20030090593 | May 15, 2003 | Xiong |
20030108238 | June 12, 2003 | Xu |
20030117611 | June 26, 2003 | Chon et al. |
20040114799 | June 17, 2004 | Xu |
20040151401 | August 5, 2004 | Sawhney et al. |
20060188013 | August 24, 2006 | Coimbra et al. |
20100182511 | July 22, 2010 | Xu |
1117251 | July 2001 | EP |
2000227960 | August 2000 | JP |
WO 9844739 | October 1998 | WO |
WO99/07156 | February 1999 | WO |
WO 00/13407 | March 2000 | WO |
WO 0158129 | August 2001 | WO |
WO 03043348 | May 2003 | WO |
- R. Tsai et al., “Multiframe Image Restoration and Registration,” Advances in Computer Vision and Image Processing, vol. 1, pp. 317-320, 1984.
- S. Borman et al., “Spatial Resolution Enhancement of Low-Resolution Image Sequences a Comprehensive Review with Directions for Future Research,” Laboratory for Image and Signal Analysis, University of Notre Dame, Jul. 1998.
- S. Park et al., “Super-Resolution Image Reconstruction: A Technical Overview,” IEEE Signal Processing Magazine, pp. 21-36, May 2003.
- C. Segall et al., “High-Resolution Images from Low-Resolution Compressed Video,” IEEE Signal Processing Magazine, pp. 37-48, May 2003.
- A. Tekalp et al., “High-Resolution Image Reconstruction from Lower-Resolution Image Sequences and Space-Varying Image Restoration,” IEEE 1992.
- R. Schultz et al., “Subpixel Motion Estimation for Super-Resolution Image Sequence Enhancement,” Journal of Visual Communication and Image Representation, vol. 9 No. 1, Mar. 1998.
- P. Eren et al., “Robust, Object-Based High-Resolution Image Reconstruction from Low-Resolution Video,” IEEE Transactions on Image Processing, vol. 6 No. 10, Oct. 1997.
- Y. Rubnew, “Empirical Evaluation of Dissimilarity Measures for Color and Texture,” Computer Vision and Image Understanding, vol. 84, 2001.
- Musmann et al., “Advances in picture coding,” Proceedings of the IEEE 73(4), 1985, 523-548.
- Bernd Heisele, “Motion-Based Object Detection and Tracking in Color Image Sequences,” Proceedings IEEE 4th International Conference on Image Processing (ICIP-97), vol. III, Oct. 1997.
- Hansen, M., et al.: “Real-Time scene stabilization and Mosaic Construction”; Image Understanding WorkShop; NDv. 1994; pp. 457-465; XPOOO852887.
- Kang-Sun Choi, et al., “An Efficient Digital Images Stabilizing Technique for Mobile Video Communications”: IEEE, Piscataway. NJ.: Jun. 15, 2000; pp. 246-247; XP002326257.
- J.-R Ohm, P. Ma : “Feature-based cluster segmentation of im˜e sequences,” Aroceedings IEEE 4th Inte{national Conference on Image Processing (ICIP-97), va .111, pp. 178-181, Oct. 1997.
- J. Y. A. Wang and E H. Adelson. Layered Representation for Motion Analysis. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition 1993. pp. 361-366, New York, Jun. 1993.
- David Tweed and Andrew Calway. Moving Object Graphs and Layer Extraction from Image sequences. In Tim Cootes and Chris Taylor, editors, Proceedings of the British Machine Vision Conference, pp. 133-142. BMVA, Sep. 2001.
- Qiard G. Medioni, Isaac Cohen, Franc;ois Brenond, Somboon Hongeng, Ramakant Nevatia: Event Detection and Analysis from Video Streams: IEEE Transactions on Pattern Analysis and Machine Intelligence 23(8): 873-889 (2001).
- Remi Megret, Daniel DeMenthon. A Survey of Spatio-Temporal Grouping Techniques. Research report CS-TR-443, LAMP, University of Maryland, Aug. 2002.
- Zhu ZhiGang, Xu GuangYou, Yang YuOong, Jin J S. Camera stabilization based on 2.50 motion estimation and inertial motion filtering[A). IEEE. International Conference on Intelligent Vehicles[C]. Stuttgart, Germany: IEEE, 1998.
- J. Y. A. Wang and E. H. Adelson. Spatio-Temporal Segmentation of Video Data. Proceedings of SPIEon Image and Video Processing II, vol. 2182, pp. 120-131, San Jose. Feb. 1994.
- M. J. Black and P. Anandan. The robust estimation of mUltiple motions: Parametric and piecewise-smooth flow fields. Computer Vision and Image Understanding, 63(1 ):7.5-104, Jan. 1996.
- Irani et al., “Improving Resolution by Image Registration,” CVGIP: Graphical Models and Image Processing, vol. 53, No. 3, pp. 231-239 (May 1991).
- Adelson et al., “Pyramid Methods in Image Processing,” RCA Engineer, 29-6, pp. 33-41 (Nov./Dec. 1984).
Type: Grant
Filed: Apr 15, 2004
Date of Patent: Oct 11, 2011
Patent Publication Number: 20050232514
Assignee: Hewlett-Packard Development Company, L.P. (Houston, TX)
Inventor: Mei Chen (Mountain View, CA)
Primary Examiner: Bhavesh Mehta
Assistant Examiner: Sean Motsinger
Application Number: 10/824,692
International Classification: G06K 9/32 (20060101);